I've been thinking lately about giving Tirelat more of a history, trying 
to figure out how it got the way it is (and potentially connecting it 
with other related languages if you go back far enough). One of the 
issues is what happened to the /ð/ phoneme which existed at least in 
some dialects. In some cases it merged with /d/ and in others with /v/. 
I figure something like a conditioned merger must have happened. But 
I've been trying to understand what happened in General American with 
the phoneme that ended up as /ɒ/ in RP, and I can't find any consistent 
features that would explain the splits and mergers. Is it possible that 
phoneme mergers and splits between dialects of a language can be 
irregular, with some words going one way and some words another, in a 
more or less arbitrary way?

It looks like most words with /ɒ/ in RP have either /ɑ/ or /ɔ/ in 
General American. In cases where GenAm has /ɔ/, the following consonant 
is often -f -g, -ŋ, -r, -s, or -θ. If the following consonant is -l, it 
can go either way (doll, follow, solid, tolerant with /ɑ/ vs. dolphin, 
golf, malt, solve with /ɔ/). But maybe it depends on whether the /l/ is 
followed by another consonant.

So if this is a conditioned merger, it's a pretty odd set of conditions. 
But even this set of rules seems to be incomplete, since there are 
apparent exceptions. You'd expect "hostage" and "hostile" to have /ɔ/ 
(like "cost" or "lost"), but they have /ɑ/. Other apparent exceptions 
that have /ɑ/ where you might expect /ɔ/ include "borrow", "cog", 
"jostle", and "wasp". There's also "sorry", which the way I say it is 
anomalous in having /ɒr/ where no other words have /ɒ/ before /r/. And 
before -ʃ, it looks like /ɒ/ can go either way (Josh with /ɑ/ vs. wash 
with /ɔ/), but there are too few examples to generalize.

Might there have been some regularity in the splits and mergers that was 
obscured by later sound changes? Or is it possible that phonemes with 
only marginal contrast can be lost in unpredictable patterns? (I can 
only think of "coral" vs. "choral" as a contrast that was lost in the 
mergers, but there may be a few others.)

I'm thinking that Tirelat /ð/ may have been an uncommon phoneme in the 
first place, so I can probably come up with a set of conditions that 
determine whether it merges with /v/ or /d/ without being too 
inconsistent with the existing vocabulary. But it would be nice to make 
sense of how these kinds of change can happen in general.